Viewpoint: The AI application layer should not be priced based on tokens, but should be anchored to "recognizable work value."

a16z partner Sarah Wang recently published an article pointing out that AI application layer products should not price based on tokens like the model layer, but rather on "recognizable work units." The article argues that token pricing anchors the value of application products to a unit whose cost is continuously declining, making it difficult for customers to predict context length, retrieval volume, or reasoning time, and improperly compares applications to raw computing power.

The article suggests a tiered pricing model based on value levels: model layer priced by tokens; application layer priced by recognizable work units for customers (such as account research briefs, code modifications, completed queries), which can be encapsulated through Credits; and scenarios that are attributable and have clear value priced directly by results (such as resolved customer service conversations, qualified leads). The design of Credits should map to different levels of work difficulty to protect gross margins and distinguish "work value" from "delivery cost." The article uses Clay as an example, where its new pricing separates Data Credits (third-party data) from Actions (orchestrated work), only passing on costs for reasoning models with significant cost fluctuations without markup. The author believes that pricing anchored to value rather than computing cost allows customers to understand spending in relation to value, while also benefiting product providers in maintaining profit margins.